AI in Banks: Fraud Desks Report Losses Avoided While Bank-Wide Programmes Still Count Hours

At a glance

• AI Atlas holds 46 published use cases in Banks from 36 institutions in 19 countries. 42 are deployments, 3 are experiments and 1 is research. The United States (10), the United Kingdom (8) and the Netherlands (4) lead.
• Fraud is the one area where banks report AI results in money lost or saved, and the only one where an agent writes the controls. At one bank the agent proposes most new card-fraud rules, and humans approve them before they go live.
• Bank-wide programmes are starting to report value in currency rather than model counts. Only a few show how the number was measured: DBS cites audited value and controlled comparisons, and Santander breaks its quarterly figure down by source.
• Customer assistants that began as question-answering bots are being rebuilt as agents that act on the account. Most of the agentic versions are still limited rollouts.
• In documents and credit, the first agents prepare a file and a human decides. No case in the set lets an AI make a lending decision on its own.

I. Fraud is where banks let the agent write the rules

Fraud is the one domain where losses are booked, so it is also the one where results can be stated plainly. Commonwealth Bank of Australia: agent that writes fraud rules goes furthest. The agent watches more than 80 million signals a day, and when it finds a new pattern it proposes a detection rule. It has developed or updated 75% of the bank's card-fraud rules. The bank's analysts approve each rule before deployment. CommBank says its fraud technology contributed to a fall of more than 20% in fraud losses in the first half of FY26 compared with a year earlier. The bank's in-house teams built the agent in three months on top of a platform that was already monitoring those signals.

The strongest third-party result is from Ireland. Bank of Ireland: Featurespace payment-fraud platform reports that since it adopted the platform in 2024, customer fraud losses fell 25%, false positives fell 87.1%, alert volumes fell 85% and case-handling time halved. The last three numbers matter as much as the first. In fraud operations the cost is not only money lost but analyst time spent on legitimate payments.

The anti-money-laundering cases are less specific. Commerzbank: Hawk AI risk layer for AML adds a model on top of the bank's existing rules engine instead of replacing it. The bank reports higher alert accuracy and fewer false positives, but gives no figures. Zürcher Kantonalbank: behavioural fraud detection describes a tiered design. Light models screen every transaction, about 95% pass straight through, and deep-learning and graph models examine the rest. The bank has been tuning the platform since 2021. Both cases share a pattern: a specialist vendor's model sits beside the rules engine rather than replacing it, which keeps each decision explainable to a supervisor.

II. Bank-wide programmes now report value in money, but few show the method

A second group of cases describes AI across the whole bank. The way banks report it is changing. Earlier cases count models and use cases; the newer ones report money.

• DBS Bank: S$1 billion in audited AI value is the only case that describes its method. DBS reports S$1 billion of audited economic value in FY2025, up 33% on the year before, from more than 1,500 models. It says it measures that value by comparing AI-supported groups against control groups.
• Santander: 280 agents in production gives the most detailed breakdown. Its $40 million return in Q1 2026 splits into about $18 million from voice agents in customer service, $11 million from faster KYC and AML processing, and $11 million from code assistance. The bank has kept its target of $1.15 billion by 2028, but warns that reaching it depends on regulators and on rollout to its Latin American subsidiaries.
• JPMorgan Chase: LLM Suite for 250,000 employees reports about $1.5 billion in cumulative savings from fraud prevention, personalisation, trading, operations and credit. It does not say how much of that comes from the generative tools the case is about.
• NatWest: £1.2 billion transformation, 70,000 hours saved and Lloyds Banking Group: agentic AI for operations sit between the two styles. NatWest reports hours saved and freed investment capacity. Lloyds reports about £50 million of value from more than 50 use cases in 2025, and projects £100 million from agentic AI in 2026. The 2026 figure is a forecast.

The contrast is between numbers that have been checked and numbers that are only announced. A bank that cites audit and control groups, or that breaks its figure down by source, can be held to it next year. A cumulative or projected figure cannot.

III. Customer assistants are moving from answering to acting

The customer-facing assistants have the longest history in the set. Bank of America: from Erica to enterprise AI dates Erica to 2018 and reports a 98% containment rate. Containment means the conversation ends without a human agent. The bank's employee version cut IT service-desk calls by more than half. Wells Fargo: Fargo virtual assistant reports more than 242 million fully automated interactions, a figure published by its cloud provider.

The newer entrants show how quickly the bar has moved. ING: generative chatbot for Dutch support went live in September 2023, when the scripted bot it replaced resolved only 40–45% of 85,000 weekly contacts. In its first seven weeks the new bot helped 20% more customers resolve their issue, and its guardrails keep it away from mortgage and investment advice. By January 2026 the neobank bunq: Finn support assistant reported that its assistant handles 97% of support and fully automates 70%, in 38 languages. bunq went from concept to production in three months. Its design uses an orchestrator that routes each query to one of three to five primary agents, which call specialist agents as tools. bunq adopted this after a central router in front of many agents became a bottleneck.

The next step is assistants that act on the account, not just answer questions. NatWest: Cora from 4 to 21 journeys describes an agentic assistant built on OpenAI models. By the end of Q1 2026 it was available to 25,000 customers, a small share of the bank's customer base. Citigroup: Citi Sky wealth assistant was announced for a phased rollout to Citigold clients from summer 2026. Both are limited or future releases. So far the agentic assistants are announced far more widely than they are used.

Contact centres show the same move on the staff side. Barclays: GenAI call summaries has summarised more than 8 million US customer calls since October 2025. Scotiabank: Scotia Intelligence reports that AI handles more than 40% of contact-centre queries. It also routes about 90% of commercial-banking emails, cutting manual work by 70%.

IV. In documents and credit, the agent prepares and a human decides

Where the work ends in a credit or onboarding decision, every case keeps a human at the final step.

TD Bank: mortgage review agent is TD's first agentic use case, in production since January. The agent ingests the purchase agreement, ID, account statements and proof of income, checks them for inconsistencies, and hands a summary to a credit adjudicator. TD says review now takes minutes instead of the roughly 15 hours a human needed. The bank chose mortgages as its first case because the inputs and outputs are well defined. Citibank: four GenAI tools in Services reports an 80% cut in the time to process a new client's onboarding documents, from a few hours to 10–15 minutes. Its client-facing chatbot is still a pilot.

Two smaller cases show the same pattern at the level of a single task. Banca Sella: automated credit-rating explanations cut the time to explain a customer's rating change to a branch from 17 minutes to 5 seconds. The rating itself is unchanged; only the explanation is automated. Apoidea Group: fine-tuned model for bank documents fine-tuned an open 7-billion-parameter vision model for tables in financial statements. On the FinTabNet benchmark it scored 81.1, against 69.9 for a small commercial model. Apoidea's product runs at more than 10 financial institutions, where a financial-spreading case now takes about 10 minutes plus under 30 minutes of staff review, instead of 4–6 hours.

The older reference point is JPMorgan Chase: COIN contract intelligence. Its reported 360,000 hours of annual legal review predates generative AI. The newer cases differ mainly in scope: they now handle unstructured files such as IDs, statements and emails, not only standard contracts.

V. The plumbing for agents is being laid, and some engineers urge caution

Three cases concern infrastructure rather than a customer-facing service.

• Payments. ING Bank: first European agentic payment in production completed, with Worldline and Mastercard on 2 June 2026, a purchase that an AI agent started and the customer approved. It was the first time merchant-side AI initiation, customer approval, authentication and issuer authorisation ran together on live European card rails. The transaction record was flagged as agent-initiated, and the issuing bank kept authorisation control.
• Chips. China Merchants Bank: DeepSeek inference on domestic chips runs DeepSeek-V4 Flash on 64 Huawei Ascend processors in production. The bank plans to package the setup as a template other Chinese banks can adopt. Its stated goal is to choose among several domestic chip types model by model.
• Model supply. HSBC: Mistral AI partnership takes a European model provider partly for data-sovereignty reasons.

The most cautious voice in the set comes from a bank's own engineers. Rabobank: engineering team's agent experiments reports that a single agent confused its tools as more were added. The supervisor agent sometimes garbled what its sub-agents returned, and low-code orchestration hid what was actually happening. The team's verdict, published at the end of September 2026, is that its agent workflows are "not yet stable and reliable enough for production". That was three months after Santander said it had 280 agents in production. The two positions need not conflict: Santander's largest categories are narrow, well-bounded tasks, which are exactly the kind Rabobank's engineers recommend starting with.

Where it is still thin

• Autonomous credit decisions. No case lets an AI approve or decline a loan. Given supervisors' model-risk rules this may be deliberate, but it means the set says nothing about what happens when a bank tries.
• Figures for AML. The money-laundering cases describe architecture and name vendors but give no detection or false-positive numbers. They lag well behind the fraud cases.
• Emerging markets. Outside China, the set has one Nigerian bank (a voice-banking launch with no usage figures) and two Latin American lenders. One Latin American case rests on a vendor partnership announcement; the other on an undated source.
• Headline figures from secondary sources. Several of the most-repeated numbers in the set, such as model counts and interactions per second, come from compilations and listicles rather than from the banks. This article leaves those out.
• Failures. As in other industries, no case describes an AI system that a bank withdrew or that caused a loss.

What to watch

1. Does another bank let an agent write production controls? If a second large bank reports agent-proposed fraud or AML rules going live under human approval before mid-2027, CommBank's model is becoming standard practice. If none does, it stays an outlier that depends on one bank's in-house platform.
2. Do agentic assistants grow past pilot size? NatWest had 25,000 customers on its agentic assistant at the end of Q1 2026. A figure in the millions by the 2026 annual results would show that banks trust agents that act on the account. A figure still in the tens of thousands would show they do not yet.
3. Does Santander's quarterly figure hold up? If the next quarters keep a similar split between voice deflection, KYC/AML and code assistance, the $1.15 billion target is on track. A shift toward projected rather than realised value would suggest the early quarters were the easy wins.
4. Do agentic payments reach recurring use? ING, Worldline and Mastercard named recurring and delegated payments as next steps. A live recurring agentic payment in Europe within a year would mean card rails are ready for agents; if not, the June transaction remains a demonstration.